Today's AI Tip
Want AI emails that sound like you? Paste three you sent.
Start with one rough note
The fictional note: Jess's logo is two days late because she added a fourth colour after sign-off on three. New date Fri 16 Oct, $150 extra round per the contract, and no guilt.
![No samples gives the template voice. With just "Turn my rough note into an email to my client," it wrote about 120 words, with a subject line, a stock opener and "[Your Name]".](https://media.carouselcdn.com/ai-tools/2026-10-05-a/03.png)
No samples gives the template voice
Real output · GPT-5.5 (OpenAI API) · Oct 2026
With just "Turn my rough note into an email to my client," it wrote about 120 words, with a subject line, a stock opener and "[Your Name]".
Source: Own test, GPT-5.5 (OpenAI API), Oct 2026
Paste three emails, then ask
Tested in GPT-5.5 (OpenAI API), Oct 2026, with three fictional emails
Both Anthropic and OpenAI recommend examples to steer tone and format. Swap in your own three emails.
Source: Anthropic and OpenAI guides, 2026
What's in Sam's three samples
The three fictional samples are short, plain, open with "Hi Name —" and sign off "Cheers, Sam". In this test the AI borrowed the opener and sign-off, but copied no full sentence.
Source: Own test, GPT-5.5 (OpenAI API), Oct 2026
Same note, now in Sam's style
Real output · GPT-5.5 (OpenAI API) · Oct 2026
With three samples, the same note became 51 words in Sam's style. The date, the $150 and the contract clause all stayed. The reason stayed too, but with a catch (next slide).
Source: Own test, GPT-5.5 (OpenAI API), Oct 2026
Check who caused the delay
In all four runs, with and without samples, it swapped "she added" for "we" ("we added" or "we're adding"). That softens the tone but changes who caused the delay. The version with samples also dropped the subject line, so add your own.
Source: Own test, GPT-5.5 (OpenAI API), Oct 2026
Keep three emails ready
Keep three short sent emails in a note and paste them above each rough draft. Remove client names, account numbers and personal details first, and check your employer's approved tools before pasting real emails.
Try it: Rewrite your next email both ways and compare.
Sources and assumptions
- Anthropic, Prompting best practices (Claude Platform Docs), 'Use examples effectively': "Examples are one of the most reliable ways to steer Claude's output format, tone, and structure. A few well-crafted examples (known as few-shot or multishot prompting) improve accuracy and consistency." It also recommends wrapping examples in <example> tags so the model can tell them from instructions. (checked 2026-10-04)
- OpenAI API docs, Prompt engineering guide, 'Few-shot learning': "Few-shot learning lets you steer a large language model toward a new task by including a handful of input/output examples in the prompt, rather than fine-tuning the model." Diverse examples work best. (checked 2026-10-04)
- Own test: engine/ai-test.js, GPT-5.5 via the OpenAI API (default settings, no system prompt), 4 runs: The real outputs in record.tests: no-samples-1/-2 (119 words with subject line and [Your Name] placeholder; a similar length with the same template) and with-samples-1/-2 (about 51 words, 'Hi Jess —', 'Cheers, Sam'). All facts kept in all 4 runs, and all 4 changed 'she added' to 'we added'. (checked 2026-10-04)
Assumptions:
- Sam, Jess, Dana, Marco and Priya and all the emails are fictional sample material written for this test.
- Tests ran on GPT-5.5 through the OpenAI API (gpt-5.5, default reasoning and temperature, user message only, no system prompt) on 2026-10-04 US time (logged 2026-10-05 UTC). Label: 'GPT-5.5 (OpenAI API), Oct 2026', not the ChatGPT app.
- Claude was NOT tested. The Claude runs in tests.json failed (no output), so the slides must not claim this works in Claude. They may cite Anthropic's docs recommending examples for tone, as documentation only.
- One note, two runs per version. That's a single small test, not proof that it works every time.
- Word counts were done by hand, not counting dashes: no-samples-1 = 119 words including the subject line and sign-off; with-samples-1 = 51 words including 'Cheers, Sam'. The earlier figure for the sample emails ('about 25–31 words') was miscounted and has been withdrawn. The 51-word output is longer than the sample emails, so the AI got much shorter but did not match their length.
- 16 Oct 2026 is a Friday, so the note's 'fri 16 oct' is internally consistent.
The short version
- Paste three emails you actually sent
- Ask it to match your tone and length
- Tell it not to copy your sentences
- Check the facts and add a subject line
